الملخص
Recent world action models (WAMs) reuse pretrained video VAEs whose encoder latents directly condition downstream action policies. Quantization must therefore preserve not only reconstruction fidelity but also the policy-facing latent contract expected by the frozen policy. Direct NVFP4 leaves W4A4 quantization error uncompensated, whereas joint quantization-aware training (QAT) can recover reconstruction by moving this representation. On Wan2.1, joint QAT nearly matches FP32 VBench-7 (0.7403 versus 0.7409), yet LIBERO success collapses from 95.5% to 10.5%. Controlled decoder-only experiments show that activation quantize-dequantize operations alter the reconstruction signal and decoder Jacobian, redirecting the gradient returned to the encoder and inducing persistent latent drift. Based on this mechanism, we introduce LatentQuant, a two-stage NVFP4 QAT framework that first aligns the quantized encoder with its high-precision counterpart, then freezes it while adapting the decoder. Across Wan2.1 and Wan2.2, LatentQuant preserves near-baseline control and high reconstruction quality, achieving 95.75% success on LIBERO and 68.8% on RoboTwin. On NVIDIA B300 GPUs, NVFP4 execution achieves 1.17x-1.26x end-to-end VAE speedups over BF16 cuDNN.
الكلمات المفتاحية
بيانات النشر
- المجلة
- غير متاح
- وصول مفتوح
- وصول مفتوح أخضر
اقتبس هذه المقالة
APA 7
Deng, Z., Chen, L., Feng, S., Duan, Z., Ye, Z., Sun, Y., Li, X., Gan, R., Wang, H., & Zhang, H. (2026). LatentQuant: Preserving the Policy-Facing Latent Contract under NVFP4 VAE Quantization. https://omanscience.com/ar/articles/latentquant-preserving-the-policy-facing-latent-contract-under-nvfp4-vae-quantization
MLA 9
Deng, Ziye, et al. "LatentQuant: Preserving the Policy-Facing Latent Contract under NVFP4 VAE Quantization." https://omanscience.com/ar/articles/latentquant-preserving-the-policy-facing-latent-contract-under-nvfp4-vae-quantization.
شيكاغو (المؤلف–التاريخ)
Deng, Ziye, Lufang Chen, Shuyu Feng, Zhenwei Duan, Zicong Ye, Yu Sun, Xiaofan Li, Ruyi Gan, Hao Wang, and Hao Zhang. 2026. "LatentQuant: Preserving the Policy-Facing Latent Contract under NVFP4 VAE Quantization." https://omanscience.com/ar/articles/latentquant-preserving-the-policy-facing-latent-contract-under-nvfp4-vae-quantization.
هارفارد
Deng, Z., Chen, L., Feng, S., Duan, Z., Ye, Z., Sun, Y., Li, X., Gan, R., Wang, H. and Zhang, H. (2026) 'LatentQuant: Preserving the Policy-Facing Latent Contract under NVFP4 VAE Quantization', Available at: https://omanscience.com/ar/articles/latentquant-preserving-the-policy-facing-latent-contract-under-nvfp4-vae-quantization.
فانكوفر
Deng Z, Chen L, Feng S, Duan Z, Ye Z, Sun Y, et al. LatentQuant: Preserving the Policy-Facing Latent Contract under NVFP4 VAE Quantization. https://omanscience.com/ar/articles/latentquant-preserving-the-policy-facing-latent-contract-under-nvfp4-vae-quantization
IEEE
Z. Deng, L. Chen, S. Feng, Z. Duan, Z. Ye, Y. Sun, X. Li, R. Gan, H. Wang, and H. Zhang, "LatentQuant: Preserving the Policy-Facing Latent Contract under NVFP4 VAE Quantization," https://omanscience.com/ar/articles/latentquant-preserving-the-policy-facing-latent-contract-under-nvfp4-vae-quantization.